Proliferation of cosine-tuning in both artificial spiking and cortical neural networks during learning
Keywords: bio-functional similarity, cosine-tuning, spiking neural network, brain-machine interface
TL;DR: In a goal-driven deep-learning spiking neural network, bio-functional similarities enabled a prediction during biological new task learning, which was validated in the motor cortex of two monkeys through brain-machine interface experiments.
Abstract: Goal-driven deep learning (DL)-based artificial neural networks (ANNs) have shown many promising bio-functional similarities with different biological neural systems, though they are less reported to model the motor neural system. Even less is known about whether goal-driven DL-based spiking neural networks (SNNs) exhibit similar bio-functional properties or predictive capabilities in the motor system. In this study, we employed the motorSRNN, a recurrent SNN inspired by the primate neural motor circuit. It successfully decoded cortical spike trains (CSTs) from the primary motor cortex (M1) of two monkeys performing a reaching task. Notably, the motorSRNN replicated bio-functional properties at population and circuit levels, closely matching those observed in biology. Moreover, motorSRNN captured and cultivated more significantly cosine-tuned neurons (SCtNs) and maintained stable proliferation during learning, suggesting that similar processes may occur in a learning biological neural network. To test this prediction, we designed a brain-machine interface (BMI) experiment in which the cortical neural networks in M1 of two monkeys learned to modulate their activities to control a new decoder in four widely spaced sessions. Our results confirmed that new task learning indeed induced the stable proliferation of SCtNs in M1. In summary, the goal-driven motorSRNN demonstrates bio-functional similarity and predictive capability, offering a framework for motor circuit modeling.
Submission Number: 68
Loading